HIVE: Hidden-Evidence Verification for Hallucination Detection in Diffusion Large Language Models

Referee: [§4.2] §4.2 (Learned Evidence Selection): the description of the selector (which chooses informative step-layer pairs from the denoising trajectory) does not state whether training of this component uses ground-truth hallucination labels drawn from the same QA benchmark splits used for final evaluation. If the selector is fit on labels from the evaluation distribution (or via hyperparameter search that leaks test information), the reported gains over the eight output-only baselines could be driven by supervised feature selection on trajectory statistics rather than by any intrinsic property of the diffusion hidden states; the two-stream representation and embeddings would then be secondary.

Authors: We appreciate the referee's concern about potential data leakage in the learned evidence selector. The selector is trained exclusively using ground-truth hallucination labels from the training splits of the QA benchmarks; the validation and test splits are strictly held out and not used for selector training, hyperparameter tuning, or any form of selection. This separation ensures that performance improvements derive from the hidden-evidence signal rather than supervised leakage on trajectory statistics. We will revise §4.2 to explicitly document the training data splits, label sources, and separation from evaluation data. revision: yes

Referee: [§5.1] §5.1 and Table 1 (Experimental Setup and Main Results): the paper reports AUROC/AUPRC numbers and claims consistent outperformance but supplies no details on baseline re-implementations, hyperparameter search procedures, exact data splits, or whether any baseline had access to trajectory information. Without these, the central performance claim cannot be verified and the ablation results (which remove components of HIVE) cannot be compared fairly to the baselines.

Authors: We agree that the original manuscript lacked sufficient experimental details for reproducibility and fair comparison. In the revision, we will expand §5.1 to provide: full descriptions of how each of the eight baselines was re-implemented (including any D-LLM-specific adaptations); the hyperparameter search procedures, ranges, and selection criteria for all methods; the precise train/validation/test splits used for each benchmark; and explicit confirmation that all baselines are strictly output-only with no access to internal denoising trajectories or hidden states. These additions will enable verification of the performance claims and proper contextualization of the ablation results. revision: yes

Referee: [Table 1] Table 1 and §5.2 (Ablations): results are presented without error bars, standard deviations across random seeds, or statistical significance tests. Given the stochastic nature of both D-LLM generation and verifier LM outputs, the headline numbers (0.9236 AUROC, 0.9537 AUPRC) and the ablation deltas require uncertainty quantification to support the claim that each component is important.

Authors: We concur that uncertainty quantification is essential given the stochasticity in D-LLM denoising and verifier outputs. We will rerun all experiments across multiple random seeds (at least five), report means and standard deviations for AUROC and AUPRC in Table 1 and the ablation tables, and add statistical significance tests (e.g., paired t-tests with p-values) comparing HIVE against baselines and ablated variants. We will update §5.2 to discuss these results and their implications for component importance. revision: yes